Prime Intellect says 2,000 agents rewrote its coding agent in Rust

Why it matters

The rewrite puts Prime Intellect's own agent and infrastructure through a demanding internal workload. Company benchmarks show lower harness startup latency and memory use, while excluding inference; they do not establish better end-to-end results for customers.

Prime Intellect, co-founded by Vincent Weisser and Johannes Hagemann, says more than 2,000 agents rewrote Prime Agent, its open-source coding agent, in Rust over more than two weeks, using more than 10,000 Prime Sandboxes and over 200 billion tokens from Prime Inference's GLM-5.3 endpoint. Published on October 9th, the release puts Prime Agent's multi-agent machinery to work on the software itself: Prime Intellect says the Rust version starts faster, uses less memory and adds Windows support.

That is a product release and a live demonstration of Prime Intellect's infrastructure thesis. Co-founder Johannes Hagemann (@johannes_hage) and Weisser have built Prime Intellect around giving organizations more control over model training and agent systems. Weisser described that ambition to TechCrunch in July as broadening access to model training beyond a small group of labs. Hagemann came to the company after work on distributed training at Aleph Alpha, a technical lineage that fits Prime Intellect's infrastructure-heavy focus. The Rust rewrite applies that focus to a coding agent designed for long-running tasks and persistent subagents.

A rewrite built around checks

Prime Agent launched in August, according to Prime Intellect's release, and the company says it has since been downloaded more than 300,000 times and processed over 8 trillion tokens. Those are company-reported totals; the announcement does not break out active users, retention or how much of that usage was commercial.

For the rewrite, Prime Intellect says a root agent divided the work into dependency-ordered tasks and coordinated planning, implementation, review and verification agents. The root agent wrote no product code, Prime Intellect says, so it could focus on task allocation and integration. Separate agents checked the work rather than relying on the same agent to write and approve its own changes.

The verification was more concrete than a claim of agent autonomy alone. Prime Intellect compared the Rust and TypeScript versions against scripted tasks, checking rendered terminal frames, session transcripts, requests sent to model providers and daemon-protocol messages. It also had agents audit features component by component. The process reduced the need for human review during the initial port, according to Prime Intellect, but it did not remove people from release decisions: the company says its engineers directed follow-up work, investigated issues found in internal use and reviewed the results.

Prime Intellect says internal testing exposed gaps the automated parity checks had not caught. It moved its own agents and staff onto the Rust version, then spent the following weeks fixing bugs, finishing feature work and polishing the interface. Passing scripted parity tests established that selected behaviors matched, not that every real-world workflow was already production-ready.

What the speed figures measure

Prime Intellect's benchmark compared Rust and TypeScript builds on fresh four-core, 8 GB sandboxes. Its hillclimb table reports cold-start input-ready latency of 51.9 milliseconds for Rust versus 736.1 milliseconds for TypeScript. For large-session process-tree RSS on a 10 MiB session, it reports 237.3 MB for Rust versus 1,130.0 MB for TypeScript. The measurements use a scripted model and exclude inference, so they speak to the agent harness's startup and runtime overhead, not how quickly a model completes an end-to-end coding task.

Prime Intellect also compared Prime Agent with other agent harnesses, but its post says there is no common benchmark standard and that those cross-product results should be treated cautiously. The benchmark is useful evidence of a performance change on Prime Intellect's chosen tests; it is not an independent comparison of agents across real customer workloads.

Prime Intellect's explanation for moving away from TypeScript centers on Prime Agent's persistent daemon, which streams model output, runs tools and coordinates agents across sessions. The company argues that Rust's native execution and compile-time checks better suit that workload. It also says the rewrite split the code into nine Rust crates and brought Windows support, session crash isolation and a more consistent daemon protocol.

Infrastructure becomes the product

Prime Intellect's activity visualization separates the Rust port from a performance-optimization phase. The company says the rewrite used more than 2,000 agents and over 200 billion tokens. Prime Intellect says the work ran on two eight-core CPU nodes, while agents sent compilation, type-checking and diffing jobs to Prime Sandboxes.

That combination links the agent release to the broader business. Prime Intellect offers compute, inference, sandboxes, training and evaluation tools; using those systems to build and tune Prime Agent gives the company an internal workload as well as a product showcase. The timing follows Prime Intellect's $130 million Series A, announced in July and led by Radical Ventures, with NVIDIA Ventures, Intel Capital and Dell Technologies Capital among the participants. TechCrunch reported the round at a $1 billion valuation.

Prime Agent remains open source, and the Rust release adds Homebrew installation alongside Windows beta support. The public GitHub repository describes it as an agent for coding and long-running autonomous work. The harder commercial test will be whether its reported gains in startup time and memory translate into better outcomes for users, and whether Prime Intellect can turn a self-hosted demonstration of its stack into durable use across its customer base.